Automated Machine Learning Model for Non-Alcoholic Fatty Liver Disease Prediction and External Cohort Validation

Area of Science:

• Hepatology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Non-alcoholic fatty liver disease (NAFLD) is a prevalent liver condition often diagnosed late.
• Current diagnostic methods like imaging and biopsy have limitations for early screening.
• There is a critical need for efficient, non-invasive NAFLD detection tools.

Purpose of the Study:

• To develop and validate an automated machine learning (AutoML) model for early non-alcoholic fatty liver disease screening.
• To identify key clinical predictors for NAFLD using advanced feature selection techniques.
• To assess the model's diagnostic accuracy and external validity in independent cohorts.

Main Methods:

• Leveraged a large NHANES dataset (n=2677) for model development and an external cohort (n=200) for validation.
• Employed a two-stage feature selection combining LASSO regression and ChatGPT-4 analysis.
• Utilized AutoML to integrate multiple algorithms, with performance evaluated by ROC curves, F1 scores, and SHAP analysis.

Main Results:

• The Gradient Boosting Machine (GBM) model achieved high AUCs: 0.843 (training), 0.851 (testing), and 0.945 (external validation).
• Key predictors identified include Body Mass Index (BMI), triglycerides, and Gamma-Glutamyl Transferase (GGT).
• SHAP analysis confirmed the significance of these variables in NAFLD prediction.

Conclusions:

• The AutoML-driven model provides a reliable, non-invasive, and efficient method for early NAFLD detection.
• This approach reduces reliance on expert judgment and enhances diagnostic precision.
• The model shows significant potential for widespread clinical application in managing NAFLD.